Home / Blog / A dataset card template for licensed AI training data

A dataset card template for licensed AI training data

A dataset card is a short document that ships with a dataset and answers: what is in it, where it came from, what the sources agreed to, how it was labelled, how personal information was removed, what the license allows, and what is known to be missing. The template below covers all of these and is free to copy.

Dataset cards grew out of research proposals such as Datasheets for Datasets (Gebru et al.). For licensed commercial data, a card also has to answer legal and provenance questions that research datasets often skip. This is the template we use at Datadhara. Copy it, adapt it, and ask your suppliers to fill it in.

The template

# Dataset card: <name> ## 1. Summary - What the data is, in one paragraph - Intended uses (training / evaluation / fine-tuning) - Uses that are out of scope ## 2. Contents - Record types and counts - File formats and schema (link to field descriptions) - Languages, scripts and code-mixing - Time period covered ## 3. Sources and consent - Number and type of source organisations (no names unless agreed) - What each source agreed to: AI training? commercial use? duration? - Whether end individuals were notified, and how - Withdrawal process ## 4. Collection - How records were captured (scan, phone photo, export, recording) - Selection criteria: what was included and excluded, and why - Collection dates ## 5. Labelling - Label set and definitions - Who labelled (in-house, vendor, domain experts) and with what training - Quality checks: double labelling, agreement rates, review sample size - How ambiguous cases were resolved ## 6. De-identification - Identifier types removed (list them) - Method: automated, manual, or both; who reviewed - Where it happened (at the source / after transfer) - Residual risks and how they were assessed ## 7. Known gaps and biases - Under-represented groups, regions, formats or conditions - Known label noise - Anything a model trained on this data is likely to get wrong ## 8. License - Field of use, exclusivity, term, region - Model ownership and output restrictions - Deletion obligations at end of term ## 9. Provenance and audit - How each record links to its source agreement (e.g. consent_ref) - Who to contact for audit requests ## 10. Version history - Version, date, what changed

How to use it

A worked example

Our free synthetic sample files show what sections 2, 6 and 9 look like in practice: pseudonymous IDs, month-level dates, five-year age bands and a consent_ref on every row. The training data rights checklist turns sections 3 to 8 into questions to ask any supplier.

Questions

What is a dataset card?

A short document shipped with a dataset describing its contents, sources, consent, collection, labelling, de-identification, license terms and known gaps.

What should a dataset card include for licensed data?

Beyond contents and labelling, it should state what each source agreed to, how personal information was removed and where, the license terms, and how every record links back to its source agreement.

Where do dataset cards come from?

They grew out of research proposals such as Datasheets for Datasets (Gebru et al.) and model and data cards used on open model hubs.

Tell us what your models are missing

A few lines is enough: the domain, the format, and roughly how much.

Request data →